ADA613693

Building Task-Oriented Applications: An Introduction to
the Legion Programming Paradigm
by Richard H Haney, Song J Park, and Dale R Shires
ARL-TR-7206 February 2015
Approved for public release; distribution is unlimited.

NOTICES
Disclaimers
The findings in this report are not to be construed as an official Department of the Army position unless
so designated by other authorized documents.
Citation of manufacturer’s or trade names does not constitute an official endorsement or approval of the
use thereof.
Destroy this report when it is no longer needed. Do not return it to the originator.

Army Research Laboratory
Aberdeen Proving Ground, MD 21005-5067
ARL-TR-7206 February 2015
Building Task-Oriented Applications: An Introduction to
the Legion Programming Paradigm
Richard H Haney
Oak Ridge Associated Universities
Song J Park and Dale R Shires
Computational and Information Sciences Directorate, ARL

ii
REPORT DOCUMENTATION PAGE Form Approved
OMB No. 0704-0188
Public reporting burden for this collection of information is estimated to average 1 hour per response, including the time for reviewing instructions, searching existing data sources, gathering
and maintaining the data needed, and completing and reviewing the collection information. Send comments regarding this burden estimate or any other aspect of this collection of information,
including suggestions for reducing the burden, to Department of Defense, Washington Headquarters Services, Directorate for Information Operations and Reports (0704-0188), 1215 Jefferson
Davis Highway, Suite 1204, Arlington, VA 22202-4302. Respondents should be aware that notwithstanding any other provision of law, no person shall be subject to any penalty for failing to
comply with a collection of information if it does not display a currently valid OMB control number.
PLEASE DO NOT RETURN YOUR FORM TO THE ABOVE ADDRESS.
1. REPORT DATE (DD-MM-YYYY)
February 2015
2. REPORT TYPE
Final
3. DATES COVERED (From - To)
July 1, 2014–October 28, 2014
4. TITLE AND SUBTITLE
Building Task-Oriented Applications: An Introduction to the Legion Programming
Paradigm
5a. CONTRACT NUMBER
5b. GRANT NUMBER
5c. PROGRAM ELEMENT NUMBER
6. AUTHOR(S)
Richard H Haney, Song J Park, and Dale R Shires
5d. PROJECT NUMBER
5e. TASK NUMBER
5f. WORK UNIT NUMBER
7. PERFORMING ORGANIZATION NAME(S) AND ADDRESS(ES)
US Army Research Laboratory
ATTN: RDRL-CIH-S
Aberdeen Proving Ground, MD 21005-5067
8. PERFORMING ORGANIZATION
REPORT NUMBER
ARL-TR-7206
9. SPONSORING/MONITORING AGENCY NAME(S) AND ADDRESS(ES) 10. SPONSOR/MONITOR’S ACRONYM(S)
11. SPONSOR/MONITOR'S REPORT
NUMBER(S)
12. DISTRIBUTION/AVAILABILITY STATEMENT
13. SUPPLEMENTARY NOTES
14. ABSTRACT
The Legion programming system from Stanford University was developed to address the specific needs of portable parallel
heterogeneous programming with emphasis on program data. Data-centric Legion abstracts the underlying hardware such that
the focus is algorithmic development rather than the idiosyncrasies of different computing architectures. Legion is part of a
small but growing movement to treat parallel programming design and development in a task-oriented fashion and, as of this
writing, is the only one to address this paradigm dynamically. This work employs a gravitational n-body simulation as a vehicle
to introduce Legion to a wider audience of parallel programmers.
15. SUBJECT TERMS
n-body, software design, Legion, task-oriented, parallel programming
16. SECURITY CLASSIFICATION OF:
17. LIMITATION
OF ABSTRACT
UU
18. NUMBER
OF PAGES
18
19a. NAME OF RESPONSIBLE PERSON
Richard H Haney
a. REPORT
Unclassified
b. ABSTRACT
Unclassified
c. THIS PAGE
Unclassified
19b. TELEPHONE NUMBER (Include area code)
410-278-7866
Standard Form 298 (Rev. 8/98)
Prescribed by ANSI Std. Z39.18

iii
Contents
List of Figures iv
List of Tables iv
1. Introduction 1
2. Background 1
2.1 The N-body Problem.......................................................................................................2
2.2 The Legion Programming Model....................................................................................3
2.3 Registering Legion Tasks................................................................................................3
2.4 Launching Legion Tasks .................................................................................................4
2.5 Domain of Execution.......................................................................................................4
2.6 Legion Is Object-Oriented...............................................................................................5
2.7 Computing Environment Employed................................................................................5
3. Building the Application 6
3.1 Defining Tasks ................................................................................................................6
3.2 Final Algorithm...............................................................................................................7
4. Observed Results 7
5. Conclusions and Future Work 8
6. References 9
Distribution List 11

iv
List of Figures
Fig. 1 Pair-wise n-body algorithm .................................................................................................2
Fig. 2 Legion task registration hierarchy.......................................................................................3
Fig. 3 Legion region descriptions ..................................................................................................4
Fig. 4 Top-level task ......................................................................................................................6
Fig. 5 Main task .............................................................................................................................7
Fig. 6 Final application design.......................................................................................................7
List of Tables
Table N-body validity for 8 bodies................................................................................................8

1
1. Introduction
Efforts to leverage the growing computational power available using parallel programming
inexorably must address the significant factor of heterogeneous computing.1–3
Multicore central
processing units (CPUs) and various hardware accelerators exacerbate a complex architectural
landscape that inevitably constrains parallel application design and implementation.4
With the
underlying hardware changes, code must often be repurposed to remain viable. Achieving high
performance on heterogeneous systems with complex memory hierarchies proves to be a major
challenge. Furthermore, as data movement dictates computational cost and power, efficiency on
future systems will require programming models with program data reasoning.
The Legion programming system from Stanford University was developed to address the specific
issues involved with parallel heterogeneous computing from a data-centric viewpoint.5
Legion
defines a methodology whereby the developer can view parallelism through a grouping of
specially designated functions called tasks, each of which will operate on some portion of the
global domain in parallel.6
This parallelism invoked by Legion is often referred to as task-
parallelism and is implicit in the design.5,6
However, to take advantage of the benefits afforded by Legion for parallel heterogeneous
computing it is necessary to understand how the programming system works. The intent of this
study is to examine the syntax and application building potential of the Legion programming
system via an emblematic computing algorithm. The n-body problem, an important part of many
physics applications,7–9
was chosen as the vehicle to examine Legion in this regard. An outline of
the rest of this work follows.
Section 2 provides the reader with a background into both the n-body simulation utilized by this
work and the Legion programming system. This section also describes the technical
specifications of the computing environment used by this study. The details of building the
n-body simulation using Legion are discussed in Section 3. Section 4 discusses the observed
results, and the final section provides a conclusion and potential future directions.
2. Background
This section will briefly provide a background into the n-body problem as well as some pertinent
information regarding the Legion programming system. Legion is a fairly flexible system that
allows developers to create initial programs in a rather facile manner but has the ability to evolve
into far more complex structures and algorithms. Therefore, the information regarding Legion

2
in this work is far from exhaustive, and the interested reader is referred to Bauer et al5,6
and
Treichler10
for more in-depth examinations of particulars beyond the implementation of the n-
body algorithm used in this work.
2.1 The N-body Problem
The n-body problem describes an approach to understand how a set of n-bodies/particles interact
with one another over time and is based on principles proposed by Sir Isaac Newton.11
The n-
body problem is a critical component to many applications but generally falls under 2 paradigms
for computational modeling: pair-wise or hierarchical tree-based solutions.12–14
This work
follows the pair-wise computational model for gravitational simulation with an asymptotic
computational complexity of ( )2
NΟ where N is the number of bodies/particles in the system15
as performance is not the intent. As illustrated in Fig. 1, the pair-wise algorithm generates this
computational cost via the body-to-body interactions calculated in lines 6–17.
Fig. 1 Pair-wise n-body algorithm
1. SET gdt TO GRAVITY * TimeStep
2. SET EPS TO 0.0001
3. FOR T <= 0 TO TotalTime
4. FOR I <= 0 TO N
5. Particle p0 <= Particles[I]
6. FOR J <= I + 1 TO N
7. Particle p1 <= Particles[J]
8. dx <= p1.x – p0.x
9. dy <= p1.y – p0.y
10. dz <= p1.z – p0.z
11. invr <= 1 / SQRT(dx*dx + dy*dy + dz*dz + EPS)
12. invr3 <= invr*invr*invr
13. force <= p1.mass*invr3
14. ax <= ax + force*dx
15. ay <= ay + force*dy
16. az <= az + force*dz
17. END FOR
18. Xnew[I] <= p0.x + gdt*p0.velX + 0.5*gdt*gdt*ax
19. Ynew[I] <= p0.y + gdt*p0.velY + 0.5*gdt*gdt*ay
20. Znew[I] <= p0.z + gdt*p0.velZ + 0.5*gdt*gdt*az
21. p0.velX <= p0.velX + gdt*ax
22. p0.velY <= p0.velY + gdt*ay
23. p0.velZ <= p0.velZ + gdt*az
24. END FOR
25. FOR I <= 0 TO N
26. Particle p <= Particles[I]
27. p.x <= Xnew[I]
28. p.y <= Ynew[I]
29. p.z <= Znew[I]
30. END FOR
31. END FOR

3
2.2 The Legion Programming Model
The Legion programming system addresses parallelism in a data-centric fashion, providing a set
of tasks to execute on defined logical regions in the code.5,6,10
These tasks are structured in a
hierarchy and will bind associated logical regions to actual physical addresses at runtime to allow
for optimal system flexibility. A top-level task is defined for every Legion program but beyond
this the developer is free to define any number of custom operations as long as they are
registered by the system prior to execution.5,10
2.3 Registering Legion Tasks
The top-level task is executed by the Legion programming system first and can be viewed as the
control point for the system. Figure 2 shows the general hierarchy of Legion task registration
whereby the top-level task is first followed by any other tasks for the application before the
invocation of the Legion start function.5,6
Fig. 2 Legion task
registration
hierarchy
The registration of Legion tasks with the runtime engine is accomplished by passing the defined
task to the template of the register-Legion-task object along with a unique integer identifier.5,6
This allows the Legion runtime to associate the defined task with the globally unique
identification value.6
Once the defined tasks are properly registered with the Legion runtime, the
system must start execution.
The Legion start function will automatically call all underlying communication libraries such as
the Message Passing Interface (MPI) to properly initialize the system.5,6,10
The start function will
look for all defined and registered Legion tasks and execute them in the predefined hierarchy
(see Fig. 2). Upon completion, Legion will automatically call underlying communication
libraries clean-up/shut-down operations returning the resulting flag (e.g., success or failure).6

4
2.4 Launching Legion Tasks
The defined and registered Legion tasks, with the exception of the top-level task, must be
explicitly invoked via the TaskLauncher object.5
TaskLauncher objects can be passed either
single or multiple arguments with the TaskArgument or ArgumentMap objects, respectively.6
However, to ensure optimal flexibility and performance, Legion follows the delayed execution
model and does not bind arguments until the TaskLauncher object is passed to the actual
execution point in the program.5,10
2.5 Domain of Execution
The domain of execution for a Legion program is defined as the Domain object such that
structured and unstructured versions are available.5,6,10
The unstructured domain is a grouping of
points in the logical region that the developer must explicitly call and initialize, whereas the
structured domain is represented by a rectangular object that is implicitly initialized.6
These
domain definitions are validated prior to execution and represent logical regions that each task
can access and manipulate as per the dictates of the task privileges and coherence settings.5,6,10
The logical regions in Legion are broken down into sets of index spaces and field spaces.5
The
index space can be views as a row in row-column descriptions and the field space as columns
such that the field space is an applied description of entries in the current region.6
These regions
can be shared by other regions or can maintain exclusivity as shown in Fig. 3.
Fig. 3 Legion region descriptions

5
2.6 Legion Is Object-Oriented
Legion supports covariant and contra-covariant bindings, meaning that the order of sub-types
(e.g., polymorphism) is preserved or could be reversed, respectively. The ability to function in
this manner ensures that Legion will execute in an object-oriented manner that any C++
programmer should be familiar with. These arguments are passed in and out of Legion tasks in a
pass-by-value manner.
Legion permits a given task to have declared side effects on declared regions of data, resulting in
the ability of logical regions to act as heaps with tasks themselves changing state. The design of
Legion allows task to manifest as imperative or functional provided maximum flexibility within
the context of a C/C++ program. However, this design means that Legion views function
pointers as analogous to global variables and therefore the passing function pointers around to
tasks is strongly discouraged.
2.7 Computing Environment Employed
The computing environment used for this work is a 64-node heterogeneous cluster consisting of
48 IBM dx360M4 nodes, each with one Intel Phi 5110P and 16 dx360M4 nodes each with one
NVIDIA Kepler K20M graphics processing unit. Each node contained dual Intel Xeon E5-2670
(Sandy Bridge) CPUs, 64 GB of memory and a Mellanox FDR-10 Infiniband host channel
adaptor.16,17
However, this work does not employ accelerators, concentrating instead on the behavior of a
single compute node, e.g., Intel Xeon E5-2670 limited to a single processing core. The
application of the Legion programming model and runtime library to a single CPU job slot
isolates the shared memory behavior, which could then be extended to the distributed parallel
environment at a later date.
Workload management for the computing environment was implemented using IBM Platform
LSF 9.1, which is specifically designed for mission-critical high-performance computing
environments.18
The operation was limited to a single, nonaccelerator, job slot by setting the
batch submit switch n to 1 and not explicitly calling any Intel Many Integrated Core Architecture
devices.18
The BLAUNCH command was employed to allow parallel operation within the shared
memory paradigm for this work.18
The next section discusses the implementation of the gravitational n-body simulation using the
Legion programming system.

6
3. Building the Application
The first step in building the n-body simulation employed by this study is to properly install the
Legion programming system itself as described by Harada.2
Legion can be deployed as either a
shared or distributed memory system; the latter requires using the GASNet library for defined
communication calls, and the reader is referred to Harada2
for further details. After downloading
and setting the Legion runtime environment variable(s), as per Harada,2
the next step is to decide
how to partition the domain as a set of Legion tasks.
3.1 Defining Tasks
Legion operates the set of tasks as a hierarchy with the top-level task the designated control
point5
; the n-body simulation actualized for this study is no exception. The top-level task for this
work employs an unstructured set of bodies/particles and must therefore be explicitly initialized
before validation. This top-level task will instantiate a logical region that will then be loaded
with data from an external file as a structure of particles using a defined inline TaskLauncher
object (see Fig. 4). Once complete, the physical region is unmapped so that control may then
pass to the next task in the hierarchy, i.e., the main task.
Fig. 4 Top-level task
The main task defined for this work controls the outer loop of the n-body algorithm shown as
line 3 in Fig. 1. Essentially the only operation this task performs is to call the last task in the
application hierarchy, i.e., the force-update task (see Fig. 5). The main task must unmap and then
remap the current physical region when calling the force-update task as this will allow Legion to
properly schedule execution independent of the enclosing task.
1. SET Particles p TO NULL // array of particle structs
2. CALL load_file_data (file) // function loads external file
3. CREATE Logical Region lr // bind logical region
4. SET ptr_t next TO 0 // base of particles
5. CREATE TaskLauncher (WRITE) // task launcher
6. FOR I <= 0 TO N
7. SET ptr TO next->value++ // next object in particle set
8. WRITE ptr (p [I])
9. END FOR
10. UNMAP REGION
11. CALL main-task

7
Fig. 5 Main task
The force-update task for this work is derived directly from lines 6–17 and 18–23 of Fig. 1 and,
as such, is exactly the same algorithm. Therefore, for the sake of brevity, it is not repeated at this
point.
3.2 Final Algorithm
Once the task partitioning strategy has been designed, the final program is ready to deploy and
can be seen in Fig. 6.
Fig. 6 Final application design
4. Observed Results
The objective of this work is a better understanding of the practical uses of the Legion
programming system with regards to an example computing application rather than performance.
As such, it is critical that the validity of the program itself be established. This section discusses
the actual observed results of the gravitational n-body application gathered using Legion against
those from a standard C++ implementation. Both of these programs are executed against the
same input files and computing environments.
Listing 3.
1. FOR T <= 0 TO TOTAL TIME
2. UNMAP REGION
3. CREATE TaskLauncher (force-update, READ-WRITE)
4. CALL force-update // sub-task
5. REMAP REGION
6. END FOR
1. CALL top-level-task
2. LOAD input data
3. WRITE particle data to PHYSICAL_REGION
4. CALL main-task
5. FOR EACH Time Step
6. CALL force-update-task
7. END FOR
8. DONE

8
The data in the Table shows the coordinate position of the C++ version of the gravitational n-
body program used in this study with that of the corresponding Legion application. These
coordinates are for an 8-body system and clearly show the correctness of the program.
Table N-body validity for 8 bodies
Particle X Y Z X-legion Y-legion Z-legion
0 1.66 9.37 6.211 1.66 9.37 6.211
1 9.19 3.94 7.59 9.19 3.94 7.59
2 9.36 0.46 0.37 9.36 0.46 0.37
3 6.70 8.34 5.66 6.70 8.34 5.66
4 5.26 9.94 9.12 5.26 9.94 9.12
5 10.00 0.00 10.00 10.00 0.00 10.00
6 8.60 8.39 8.77 8.60 8.39 8.77
7 0.00 10.00 0.00 0.00 10.00 0.00
5. Conclusions and Future Work
This work has presented the Legion programming paradigm within the context of an important
computational algorithm employed in many of today’s high performance computing research, the
n-body simulation. The gravitational n-body algorithm was defined using the Legion
programming paradigm to illustrate the potential for future research in computing and has been
shown to be both practical and correct for this purpose. The Legion programming system is
flexible and introduces the developer to task-oriented programming, free from underlying
computing architectures and communication procedures.
Future directions for this programming model are numerous and include a more effective and
portable design for parallel heterogeneous computing. In particular the extension of this work to
include the Fast Multipole Method (FMM) of the n-body problem is an interesting direction.
FMM is currently being planned in concert with Stanford University as a part of a larger strategy
of task-oriented solutions to heterogeneous parallel computing.

9
6. References
1. Topcuoglu H, Hariri S, Wu MY. Performance-effective and low-complexity task scheduling
for heterogeneous computing. IEEE Transactions on Parallel and Distributed Systems.
2002;13(3):260–274.
2. Harada T. Heterogeneous particle-based simulation. SIGGRAPH Asia 2011 Sketches.
Proceedings of The International Conference on Computer Graphics and Interactive
Techniques; 2011 Aug 7–11; Vancouver, Canada. New York (NY): ACM; c2011. Article
19.
3. Lashuk I, Chandramowlishwaran A, Langston H, Nguyen TA, Sampath R, Shringarpure A,
Vuduc R, Ying L, Zorin D, Biros G. A massively parallel adaptive fast-multipole method on
heterogeneous architectures. In: Proceedings of the 2009 ACM/IEEE International
Conference for High Performance Computing Networking, Storage and Analysis; 2009 Nov
14–19; Portland, OR; New York (NY): ACM; c2009. Article No. 58.
4. Phothilimthana P, Ansel MJ, Ragan-Kelley J, Amarasinghe S. Portable performance on
heterogeneous architectures. In: Proceedings of the Eighteenth International Conference on
Architectural Support for Programming Languages and Operating Systems; 2013 Mar 16–20;
Houston, TX. New York (NY): ACM; c2013. p. 431–444.
5. Bauer M, Treichler S, Slaughter E, Aiken A. Legion: expressing locality and independence
with logical regions. Proceedings of SC '12 - The International Conference for High
Performance Computing, Networking, Storage and Analysis; 2012 Nov 10–16; Salt Lake
City, UT. Los Alamitos (CA): IEEE Computer Society Press; c2012. Article No. 66.
6. Bauer M, Treichler S, Slaughter E, Aiken A, McCormick P. Legion programming system.
[accessed 2014 Jun 15]. http://legion.stanford.edu/. GitHub 2014 Jan 6. [Online].
7. Fitch BG, Rayshubskiy A, Eleftheriou M, Ward T, Giampapa M, Pitman MC, Germain R.
Blue matter: approaching the limits of concurrency for classical molecular dynamics.
Proceedings of the 2006 ACM/IEEE, Conference on Supercomputing; 2006 Nov 11–17;
Tampa, FL. New York (NY): ACM; c2006. Article No. 87.
8. O’Shea BW, Bryan G, Bordner J, Norman ML, Abel T, Harkness R, Kritsuk A. Introducing
Enzo, an AMR cosmology application, in adaptive mesh refinement - theory and
applications. Chicago (IL): Springer Berlin Heidelberg; c2005. p. 341–349. Vol. 41.
9. Salmon J. Parallel N log N N-body algorithms and applications to astrophysics. In: Compcon
Spring '91. Digest of Papers; 1991, San Francisco, CA.

10
10. Treichler S, Bauer M, Aiken A. Language support for dynamic, hierarchical data
partitioning. In: Proceedings of the 2013 ACM SIGPLAN international conference on object
oriented programming systems languages and applications; 2013 Indianapolis, IN. New
York (NY): ACM; c2013. p. 495–514
11. Trenti M, Hut P. N-body simulations. Scholarpedia. 2008;3(5):3930.
12. Board J, Schulten K. The fast multipole algorithm. Computing in Science and Engineering.
2000;76–79.
13. Nyland L, Harris M, Prins J. Fast N-Body Simulation with CUDA. GPU GEMS. 2007;3(1):
677–696.
14. Grama AY, Kumar V, Sameh A. Scalable parallel formulations of the barnes-hut method for
n-body simulations. In: Supercomputing '94. Proceedings of the 1994 ACM/IEEE
Conference on Supercomputing; 1994 Nov 14–18; Washington, DC. Los Alamitos (CA):
IEEE Computers Society Press; c1994. p. 439–448.
15. Reif JH, Tate SR. The complexity of N-body simulation. In: Automata, languages and
programming. Lund (Sweden): Springer Berlin Heidelberg; 2005. p. 162–176.
16. Intel Corporation, Intel Xeon Processor E5-1600 / E5-2600/E5-4600 Product Families
Datasheet - Volume 1 and 2, Intel, 2012.
17. Intel Corporation. Intel Xeon Phi Coprocessor Instruction Set Architecture Reference
Manual, 7 Sep 2012. [accessed 2014 Jun 1]. https://software.intel.com/en-us/mic-developer.
18. IBM Corporation. IBM Platform LSF V9.1.2 documentation, [accessed 2014 Jun].
1992–2013 Jan 1 [Online]. Available: http://www-01.ibm.com/support/knowledgecenter
/#!/SSETD4_9.1.2/lsf_welcome.html.

11
1 DEFENSE TECHNICAL
(PDF) INFORMATION CTR
DTIC OCA
2 DIRECTOR
(PDF) US ARMY RESEARCH LAB
RDRL CIO LL
IMAL HRA MAIL & RECORDS MGMT
1 GOVT PRINTG OFC
(PDF) A MALHOTRA
1 DIR USARL
(PDF) RDRL CIH S
R HANEY

ADA613693

Recommended

Recommended

More Related Content

Similar to ADA613693

Similar to ADA613693 (20)

ADA613693